EfficientHRNet

EfficientHRNet is a family of scalable and efficient networks created by unifiying EfficientNet and HigherHRNet for Multi-person human pose estimation. A preprint of our paper can be found here.

Our code is based on the

Official implementation of HigherHRNet
PyTorch implementation of EfficientNet

We provide a formulation for jointly scaling our backbone EfficientNet below the baseline B0 and the rest of EfficientHRNet with it. Ultimately, we are able to create a family of highly accurate and efficient 2D human pose estimators that is flexible enough to provide lightweight solution for a variety of application and device requirements.

Environment Setup

Pytorch >= 1.1.0
Dependencies as listed in requirements.txt
COCO Keypoints Dataset along with COCOAPI as given in https://cocodataset.org/#download and https://github.com/cocodataset/cocoapi

The code was developed using python 3.6 and NVIDIA GPUs, both of which are required.

Configurations for EfficientHRNet models

Config files are found at experiments/coco/higher_hrnet/

Varying the following parameters provide different EfficientHRNet models ranging from H0 to H-4 :

scale_factor
input_size
width_mult
depth_mult

More details on scaling can be found in our paper: https://arxiv.org/pdf/2007.08090.pdf

Examples can be seen in the Example Configs folder.

EfficientHRNet Training and Evaluation

Distributed training is supported. Config settings can be customized based on user requirements. Training and validation scripts can be found at tools/

Training on COCO (Nvidia GPUs)

Single GPU training example:

CUDA_VISIBLE_DEVICES=0 python3 tools/dist_train.py --cfg experiments/coco/higher-hrnet/config.yaml

Distributed training example:

CUDA_VISIBLE_DEVICES=0,1 python3 tools/dist_train.py --cfg experiments/coco/higher-hrnet/config.yaml --dist-url tcp://127.0.0.1:12345

Testing on COCO (Nvidia GPUs)

Both single-scale and multi-scale testing are supported.

Single scale testing:

python3 tools/valid.py --cfg experiments/coco/higher-hrnet/config.yaml TEST.MODEL_FILE /path/to/model.pth

Multi-scale testing:

python3 tools/valid.py --cfg experiments/coco/higher-hrnet/config.yaml TEST.MODEL_FILE /path/to/model.pth TEST.SCALE_FACTOR [0.5,1.0,1.5]

Pretrained Models

COCO17 pretrained models for EfficientHRNet H0 to H-4 can be download here.

Method	Input Size	Parameters	FLOPs	AP	AP_multi-scale
H₀	512	23.3M	25.6B	64.0	67.1
H_-1	480	16M	14.2B	59.1	62.3
H_-2	448	10.3M	7.7B	52.8	55.0
H_-3	416	6.9M	4.2B	44.5	45.5
H_-4	384	3.7M	2.1B	35.5	39.7

Compact EfficientNet ImageNet trained weights can be downloaded here.

			ImageNet		Cifar-100
Method	Input Size	FLOPs	Parameters	Top-1	Parameters	Top-1
B0	512	0.4B	5.3M	75	4.1M	81.9
B_-1	480	0.3B	4.5M	73.8	3.5M	81.4
B_-2	448	0.2B	3.4M	71.3	2.5M	79.8
B_-3	416	0.1B	2.8M	68.5	1.9M	78.2
B_-4	384	0.05B	1.3M	65.6	1.3M	74.3

Citation

If you would like use EfficientHRNet in your work, please use the following citation.

@misc{neff2020efficienthrnet,
      title={EfficientHRNet: Efficient Scaling for Lightweight High-Resolution Multi-Person Pose Estimation}, 
      author={Christopher Neff and Aneri Sheth and Steven Furgurson and Hamed Tabkhi},
      year={2020},
      eprint={2007.08090},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

We also recommend citing EfficientNet and HigherHRNet, which inspired this work.

dhassault/EfficientHRNet